From: opencoti
Subject: [PATCH 0087] DCA: wire --dca-yarn-factor into build_dca_rope mscale (bug-740 / #550)

--dca-yarn-factor was fully plumbed (arg.cpp -> params -> cparams.dca_yarn_factor,
public C API, TS adapter) but never READ by the rope path, so the flag was inert
(bug-740). build_dca_rope now composes it into the attention scaling:
  dca_attn_factor = attn_factor * cparams.dca_yarn_factor
and passes that to both ggml_rope_multi (MROPE) and ggml_rope_ext (NEOX) in place of
the bare attn_factor. This is the YaRN mscale piece of the DCA+YaRN composition: DCA
already keeps q-k distances in-distribution via the per-regime chunk remap, so only the
softmax temperature (mscale) needs stretching for ctx past the chunk grid -- freq_scale
/ freq_base stay the model's baked values. Host-only (graph construction); the existing
CUDA rope kernel reads attn_factor from op_params unchanged, so no DSO rebuild.

Default 1.0 => dca_attn_factor == attn_factor exactly (x1.0f is bit-identity for finite
floats), so the flag-unset path is byte-identical to upstream -- the fix only makes a
previously-inert flag live; it cannot regress default behavior.

Empirical characterization (Qwen3-4B, native 40960, DCA chunk 16384, RULER 5-needle
substring under ignore_eos; NOT greedy-first-token):
  f16 KV @ 2.4x native: yarn 1.0->5/5  1.25->4/5  1.5->1/5  2.0->0/5  (monotonic: flag live)
  q8_0 KV ladder yarn=1.0: 2.4x->4  4.0x->5  5.0x->3  6.0x->4  (plain-DCA holds deep)
  q8_0 KV @ 6.0x native: yarn 1.0->4  1.1->4  1.25->4  (flat; 1.18 = principled mscale, no win)
Verdict: the mscale lever has NO beneficial regime for DCA retrieval -- it only
over-sharpens already-concentrated attention. DEFAULT 1.0 IS CORRECT. Ships as a
power-user knob (raising it degrades retrieval); documented in docs/evaluations/context.md.

diff --git a/llama.cpp/src/dca.cpp b/llama.cpp/src/dca.cpp
--- a/llama.cpp/src/dca.cpp
+++ b/llama.cpp/src/dca.cpp
@@ -164,15 +164,23 @@
 // Rope the DCA per-regime Q (or insert-K) faithfully for the arch. `pos` must be the width
 // build_attn_inp_dca allocated: 4*n_tokens for MROPE (section split below), n_tokens for NEOX.
 ggml_tensor * llm_graph_context::build_dca_rope(ggml_tensor * x, ggml_tensor * pos, const dca_rope & rp) const {
+    // opencoti-hook: dca-yarn-factor (bug-740 / #550) — compose a YaRN mscale stretch with DCA.
+    // cparams.dca_yarn_factor (1.0 = none; set by --dca-yarn-factor) multiplies the attention scaling
+    // (mscale / attn_factor) so the extending-regime softmax is re-tempered for ctx past the chunk grid
+    // (the final 2-4x of the DCA+YaRN sub-native extension recipe; see docs/features/gemma4_dca.md). DCA
+    // already keeps q-k distances in-distribution via the per-regime chunk remap, so only the mscale needs
+    // stretching — freq_scale/freq_base stay the model's baked values. Default 1.0 => this rope call is
+    // byte-identical to upstream (inert unless the flag is set).
+    const float dca_attn_factor = attn_factor * cparams.dca_yarn_factor;
     if (dca_is_mrope(rope_type)) {
         int sections[GGML_MROPE_SECTIONS] = { rp.sections[0], rp.sections[1], rp.sections[2], rp.sections[3] };
         return ggml_rope_multi(ctx0, x, pos, rp.freq_factors,
                 rp.n_rot, sections, rope_type, n_ctx_orig, rp.freq_base, rp.freq_scale,
-                ext_factor, attn_factor, beta_fast, beta_slow);
+                ext_factor, dca_attn_factor, beta_fast, beta_slow);
     }
     return ggml_rope_ext(ctx0, x, pos, rp.freq_factors,
             rp.n_rot, rope_type, n_ctx_orig, rp.freq_base, rp.freq_scale,
-            ext_factor, attn_factor, beta_fast, beta_slow);
+            ext_factor, dca_attn_factor, beta_fast, beta_slow);
 }
 
 // opencoti F5 dca #444 (DCA all-KV C1): dequant-on-lift a cached K/V tensor to f16 for the f16-only